WO2015011395A1

WO2015011395A1 - Method of estimation on a curve of a relevant point for the detection of an anomaly of a motor and data processing system for the implementation thereof

Info

Publication number: WO2015011395A1
Application number: PCT/FR2014/051882
Authority: WO
Inventors: Tsirizo RABENORO; Jérôme Henri Noël LACAILLE
Original assignee: Snecma
Priority date: 2013-07-23
Filing date: 2014-07-21
Publication date: 2015-01-29
Also published as: CA2918215C; CA2918215A1; US20160163132A1; RU2667794C2; BR112016001482B1; EP3025205B1; CN105408828B; FR3009021B1; US9792741B2; CN105408828A; EP3025205A1; BR112016001482A2; RU2016105851A; RU2016105851A3; FR3009021A1

Abstract

The invention relates to a method of estimation on a curve of a relevant point for a detection of an anomaly of a motor, said curve representing an evolution as a function of time of physical operating parameters of the motor measured by sensors on said motor, implemented by a computer linked to first means of storage, said first means of storage storing at least one profile comprising a binary code, each component of which codes a direction of variation between two consecutive characteristic points of at least one learning curve, a model making it possible to estimate a relevant point on the basis of a set of characteristic points of a curve and a filter, said method comprising: a/ (F1) the selecting of a profile stored in the first means of storage; b/ (F2) the application of the filter of the profile selected to said curve; c/ (F3) the determination of a set of characteristic points of said filtered curve and of a binary code, each component of which codes the direction of variation of two consecutive characteristic points belonging to said set of characteristic points; d/ (F4) the comparing of the determined code and of the code of the profile selected; e/ (F5) as a function of said comparison, the estimating of the relevant point on said curve on the basis of the characteristic points of said filtered curve and of the model of the profile selected.

Description

METHOD OF ESTIMATING A CURVE OF A RELEVANT POINT FOR THE DETECTION OF ANOMALY OF AN ENGINE AND A DATA PROCESSING SYSTEM FOR ITS IMPLEMENTATION FIELD OF THE INVENTION

The invention generally relates to the field of monitoring the operating state of an engine.

The invention more specifically relates to a method of estimating on a curve of a point relevant for the detection of abnormality of an engine as well as the data processing system for its implementation.

STATE OF THE ART

In the field of aeronautics, it is important to be able to monitor the operating status of an aircraft engine in order to plan and plan maintenance operations on this engine. The operating state of the engine can indeed change over time and appropriate monitoring can detect malfunctions, monitor engine degradation and plan ahead of any maintenance. This makes it possible to avoid delays on the flights, to carry out repairs before the degradation is too important, to regroup repair operations between them, etc.

Monitoring tools have been developed to identify engine-altering faults based on physical parameter measurements describing the condition of the engine.

Methods have also been described for measuring operating parameters of the engine to be monitored, calculating indicators representative of the operating state of the engine, and identifying operating anomalies of the engine from said indicators. An example of such methods is described in patents FR2939924 and FR2942001. These indicators are defined by engine behavior experts. Thus, as regards the detection of abnormalities preventing the starting of the engine, indicators such as opening time of the start valve, a time to reach a maximum acceleration of the high pressure compressor, times t1, t2 and t3 of the first, second and third start-up phase, an ignition time of the engine, a closing time of the starting valve, a gradient of the temperature of the exhaust gas or a stopping time of the engine can be defined . FIG. 1 shows the evolution of the speed 1 of the high-pressure compressor, the exhaust gas temperature 2 EGT (Exhaust Gas Temperature), the fuel flow rate 3 sent to the injectors and the pressure 4, as well as the durations t1, t2 and t3.

Such indicators can be calculated according to relevant times on the curves of the measurements of the engine operating parameters. Such relevant moments are identified on these curves by the experts.

These methods have the disadvantage of systematically involving experts to identify such times and thus make it necessary to store all of these measurement curves as long as such a location has not been achieved by a user. expert.

In order to overcome these drawbacks, tools making it possible to perform automatic detection of such relevant moments, without calling on an expert, have been developed. The use of such tools makes it possible in particular to no longer have to store a large amount of data over long periods of time, only the indicators calculated from the relevant instants determined automatically being stored in fine.

Some of these tools can in particular extract a particular relevant moment from the descriptions of this moment provided by experts during the development of the tool. Nevertheless, such solutions require developing a different tool for each type of moment relevant to detect. They also have the disadvantage of requiring the expert to finely describe the characteristics of the relevant time, in a manner understandable to the designer of the tool, so that it retranscribes these characteristics in algorithmic form.

In order to overcome these constraints, generic tools have been developed, making it possible to detect a relevant moment on any type of curve without having to adapt the tool and requiring no detailed description of the characteristics of such a moment. by an expert. Existing tools of this type may for example be based on pattern recognition. Their principle is to recognize on a curve a known characteristic form which is taken by the curve in the vicinity of a relevant moment to be detected, as represented in FIG. 2 illustrating the case where it is sought to determine the location on a curve of a characteristic shape such as the form 5.

The detection time by such tools of a relevant point on a curve is nevertheless particularly long. Indeed the entire curve is traversed to extract forms 6, 7 and 8 taken locally by the curve, and these shapes are then compared to the desired shape. In addition, such an analysis must be carried out at different scales in order to detect the desired shape in the curve irrespective of the scale at which this shape appears in the curve. Such treatments impose a large amount of calculations that slow down the detection of a relevant point on the curve.

Moreover, such tools are focused on the detection of a particular shape in the vicinity of the point of interest and neglect the information carried by the overall shape of the curve.

There is therefore a need for a generic tool for quickly detecting a relevant point on a curve without performing an expensive multi-scale analysis, while taking into account the whole of said curve and limiting the amount of data to store. SUMMARY OF THE INVENTION

The present invention thus relates, according to a first aspect, to a method of estimation on a curve of a point that is relevant for an anomaly detection of an engine, said curve representing an evolution as a function of time of physical parameters of operation of the engine. engine measured by at least one sensor on said engine,

implemented by a computer connected to first storage means,

said first storage means storing at least one profile comprising a binary code each component of which encodes a direction of variation between two consecutive characteristic points of at least one learning curve, a model for estimating a relevant point from a set of characteristic points of a curve and a filter, said method comprising:

a / selecting a stored profile in the first storage means;

bl applying the filter of the selected profile to said curve;

c / determining a set of characteristic points of said filtered curve and a binary code of which each component codes the direction of variation of two consecutive characteristic points belonging to said set of characteristic points;

Comparing the determined code and the code of the selected profile;

e / according to said comparison, estimating the relevant point on said curve from the characteristic points of said filtered curve and the model of the selected profile. Such a method makes it possible to quickly determine a relevant point on a curve, whatever the type of relevant point, without calling on an expert. Such a method also makes it possible to take account of the entire shape of the curve, while minimizing the amount of information to be stored by storing only the characteristic points of this curve. According to other advantageous and nonlimiting features:

if the determined code is different from the code of the selected profile, a new profile stored in the first storage means can be selected and the computer can execute again steps b1 to el of the method according to the first aspect;

the relevant point can be chosen from a moment of opening of a valve, a moment of net variation of a temperature or of a pressure, a moment of reaching a certain speed by a high-pressure compressor or a low pressure compressor, a moment of disengagement of a starter;

the characteristic points of curves can be chosen from inflection points, local extrema, abrupt changes of slopes; These points are particular points of a curve which make it possible to characterize the overall shape of the curve since all the curves of the same parameter measured during the same phase of operation on different engines have the same overall shape and show the same characteristic points.

- A profile may further include a threshold and the characteristic points may be consecutive local extrema whose difference ordinates is greater than said threshold;

This makes it possible to minimize the number of characteristic points to memorize while retaining only the points that are really distinct from their neighbors. - models can be generalized linear models with variable selection;

- models can check the equation: t = AX

where - 1 is the abscissa of the relevant point to estimate,

A is a line vector containing regression coefficients,

X is a column vector whose components are abscissae of the characteristic points and their transforms;

Such a model makes it possible to determine the abscissa of a relevant point solely from the abscissas of the characteristic points, without requiring a large amount of calculation.

in a mode of implementation of the method according to the first aspect, the calculator can execute: a step of estimating from the estimated relevant points of specific indicators chosen for their representativity of the operating state of the engine;

o an engine diagnostic step based on the specific indicators estimated;

o a storage step in storage means, flight after flight, specific indicators estimated and a prognostic step of a degradation of the operating state of the engine from the evolution of the specific indicators stored;

This makes it possible to automatically diagnose and predict possible engine failure or degradation of its operating state. each profile memorized in the first storage means can be determined by a learning process; this learning process for a profile can include:

a / display by a display device of several learning curves;

b / the determination by said computer of a relevant point for the anomaly detection on each of the learning curves, said relevant point of each of the learning curves being selected by an expert using input means;

c / storage in second storage means of each of said learning curves associated with said determined relevant point;

d / the selection by said calculator of a filter and a model;

e / the application by said computer of the selected filter to each of the learning curves;

f / the determination by said calculator of the characteristic points of each of the filtered learning curves;

g / the determination by said computer of the determined characteristic points of recurrent characteristic points appearing on each filtered learning curve and of a binary code whose each component codes the direction of variation between two consecutive recurring characteristic points;

h / from the determined recurring characteristic points and from the selected model, the estimation by said calculator of the relevant point; M estimation by said calculator of the error associated with the filter and model selected in step d / by comparison of the estimated relevant point with the relevant point selected by the expert in step b /;

j / storing, in said first storage means, a profile comprising said determined binary code, said filter and said model making it possible to minimize the estimated error or to obtain an estimated error lower than a determined threshold; Such a process makes it possible to characterize a relevant point automatically, without requiring an expert to precisely describe the characteristics of such a point and without requiring that a particular algorithm be defined for each type of relevant point.

According to a second aspect, the invention relates to a computer program product comprising program code instructions for executing the steps of the method according to the first aspect when said program is executed on a computer;

According to a third aspect, the invention relates to a data processing system comprising a computer, input means, at least one display device characterized in that it is configured to implement the steps of the method according to the first aspect.

Such system and computer program product have the same advantages as those mentioned for the method according to the first aspect.

PRESENTATION OF THE DRAWINGS

Other features and advantages will become apparent from the description which follows, which is purely illustrative and nonlimiting and may be read with reference to the appended figures, among which:

FIG. 1 represents a diagram illustrating the construction of specific indicators,

FIG. 2 represents a diagram illustrating an example of an algorithm based on shape recognition,

FIG. 3 schematically represents the hardware means implemented in the context of the invention,

FIG. 4 represents an exemplary graphical interface displayed to an expert in the context of the invention, FIG. 5 represents a flowchart illustrating steps of the learning process according to one embodiment of the invention,

FIG. 6 represents a diagram illustrating an example of a model making it possible to determine the abscissa of a relevant point according to one embodiment of the invention,

FIG. 7 illustrates examples of curve characteristic points,

FIG. 8 represents a flowchart illustrating steps of the estimation method on a curve of a point that is relevant for the detection of anomaly of an engine according to one embodiment of the invention.

DETAILED DESCRIPTION OF AN EMBODIMENT With reference to FIG. 3, an implementation for the invention concerns a method of estimation on a curve of a point that is relevant for an anomaly detection of an engine 9, said curve representing an evolution as a function of time of physical engine operating parameters measured by at least one sensor 10 on said engine 9.

Such a method is implemented by a computer 1 1 comprising calculation means 12, a memory 13 and a communication interface 14. This interface may allow the computer to communicate with sensors 10 able to acquire measurements of operating parameters. motor at different times. Such an interface can be a wired interface of Ethernet, USB, FireWire, serial or parallel type or a Wi-Fi or Bluetooth wireless interface.

The estimation by the computer 1 1 of a relevant point on a curve is made from comparisons to curves models called profiles. The analyzed curves are not always similar, the process uses several profiles. For example, the analysis of the temperature of Outlet gas can use two profiles, one for cold starts and one for hot starts.

Such profiles can be determined by a learning process and stored in first storage means 15. These first storage means can be in the form of a device external to the computer such as a USB external hard disk or a hard disk. Network ("NAS"). The first storage means then communicate with the computer via a communication interface such as the communication interface 14. As a variant, the first storage means can be integrated into the computer 1 1. The storage of the profiles in the first storage means can take the form of a database stored in the first storage means.

The said learning process may involve an aircraft engine operation expert to select a relevant point on learning curves. For this, the expert has a graphical interface 16 such as that shown in Figure 4, calculated by the computer 1 1. Such a graphical interface has a curve 17 in a selection window 18. This interface is displayed on a display device 19 which can be any type of screen such as an LCD, plasma, OLED or screen video projector coupled to a video projector. Such a display device is connected to the computer 1 1 by an analog or digital video connection such as Scart, VGA, DVI, DisplayPort or HDMI connection. The expert uses input means 20 to select a relevant point on a learning curve 17 displayed in the selection window 18. Such input means may consist of a keyboard and a mouse, a trackpad, a trackball or any other pointing means allowing it to specify a point on the curve 17 such as a motion detection interface. The data stored during the learning process can be stored on second means of storage 21 similar to the first storage means 15 and similarly connected to the computer 1 1.

As shown in FIG. 5, said learning process comprises the steps E1 to E1 1 described below.

During a first step E1, several learning curves Ua can be presented to the expert. These curves are displayed in a selection window 18 of the graphical interface 16. These curves are all similar although not identical and all correspond to the same type of curve on which a relevant point must be able to be determined automatically. These curves can for example be curves of temperature, pressure, air flow or fuel measured at different points of an aircraft engine or else rotational speed curves of different rotating elements of such a device. engine as high and low pressure compressors. A preselection of such a batch of similar curves may have been carried out automatically by a machine or manually by an expert from a set of curves measured on one or more aircraft engines, for example by selecting a type of size and separating curves measured over an entire flight cycle and those measured only during the start-up phase.

During a second step E2, an expert selects on each of these curves TJa a relevant point P using the input means 20. A relevant point may correspond to a particular moment of the curve such as a moment of opening of a valve, a time of net change of a temperature or pressure, a moment of reaching a certain speed by the high pressure compressor or the low pressure compressor, a moment of disengagement of the starter. The determination of such instants can make it possible to calculate specific indicators useful for estimating the operating state of an engine such as the duration of the different engine start-up phases, the ignition time, the stopping time or the maximum and average gradients of exhaust gas temperature.

In a third step E3, the learning curves T¾ are stored with their relevant point selected by the expert in the second storage means 21. The selected relevant point can be stored as its abscissa on the learning curve.

In a fourth step E4, a filter F and a model M are selected. A filter consists of a filtering function adapted to modify a curve so as to simplify the detection thereon of a characteristic point. Such filtering may consist of a smoothing, a simple or double derivation or even a treatment of accentuation of the irregularities of the curve. The corresponding filtering functions can be a Gaussian, square-shaped, triangle, Haar or Daubechies wavelet distribution.

As shown in Figure 6, a model consists of a function for determining the abscissa of a relevant point 22 from the abscissa of characteristic points 23 of a curve. Such characteristic points may correspond to local extrema, points of inflection or points of abrupt change of slope. In one variant, the model can be a generalized linear model with selection of variables. Such a model can verify the equation: t = AX

where - 1 is the abscissa of the relevant point to estimate;

- A is a line vector containing regression coefficients

X is a column vector whose components are abscissae x characteristic points and transforms of these abscissae such that In x, tan x, 1 / x ... The selection of a filter F and a model M can be carried out automatically by the computer 11, possibly randomly from a filter base and ranges of possible values of the regression coefficients, or such a selection can involve the 'expert.

In a fifth step E5 the filter F selected in the fourth step E4 is applied to each of the learning curves TJa. The application of the filter may consist of a convolutional calculation between each curve and the filter function of the filter so as to obtain filtered learning curves as shown in FIG.

In a sixth step E6, the computer 11 determines the characteristic points of each of the filtered learning curves. As shown in FIG. 7 and previously mentioned, these characteristic points may correspond to local extrema, points of inflection, that is to say having a maximum of the first derivative, or points of abrupt change of slope. that is, having a maximum second derivative between a local extremum and the point of inflection, and between the inflection point and the other local extremum. In the case of the determination of local extrema, it is possible to minimize the number of points retained by retaining at the end of this determination only consecutive local extrema whose difference in ordinate is greater than a first predetermined threshold. Advantageously only the abscissa of these characteristic points is memorized.

During a seventh step E7, the calculator determines, among all the characteristic points of the filtered learning curves TJa, recurring characteristic points. These recurring points are characteristic points detected in the majority of filtered filtered learning curves. According to one variant, these recurring characteristic points are determined solely from the local extrema of learning curves. In this variant, the recurring characteristic points other than the consecutive recurrent local extrema are determined a posteriori and in the following manner: a point of inflection is chosen between two consecutive local extrema, and if there are several, we choose the one having the maximum ordinate on the first derivative. A point representing a sudden change in variation is chosen between a local extremum and a point of inflection, and a point of inflection and a local extremum. If there are several points of sudden variation, one chooses the one having the maximum ordinate on the second derivative.

The calculator 11 also determines a binary code C, each component of which codes the direction of variation between two consecutive recurring characteristic points. For example a "1" can encode the fact that a characteristic point of ordinate y1 is followed by a characteristic point of ordinate y2 greater than y1, and a "0" can then code the fact that a characteristic point of ordinate y1 is followed by a characteristic point of ordinate y2 lower than y1. Such a code then constitutes a binary representation of the ordinate profile of the recurring characteristic points common to the majority of the learning curves Ua. The filtered learning curves on which the identified recurring characteristic points do not appear may be set aside and may be used to determine another profile during a subsequent learning process. A first profile is thus determined from a maximum of learning curves and a second profile is determined from a maximum of curves among the remaining curves ...

During an eighth step E8 the computer 1 1 determines the abscissa of a relevant point P 'on one or more of the learning curves Ua from the recurring characteristic points, in particular their abscissae, determined in the seventh step E7 and the model M chosen in the fourth step E4. In a ninth step E9, the calculator 11 estimates an error in determining each of the relevant points determined at the eighth step E8 by comparing the abscissa of a relevant point P 'determined at the eighth step E8 and the abscissa. of the relevant point P selected by the expert on the same learning curve in the second step E2. The calculator then determines the root mean square of all the estimated determination errors. This average determination error is associated with the filter F and the model M selected in the fourth step E4.

In a tenth step E10, the calculator determines whether the average error for determining the relevant point estimated at the ninth step E9 is sufficiently small to consider the determination of the relevant points made at the eighth step E8 as satisfactory. According to one embodiment, the computer compares the average error of determination with a second predetermined threshold. As long as the average error of determination is greater than this second predetermined threshold, the computer rejects the filter F and the model M selected in step E4, selects a new filter and a new model, and then again implements the steps E5 to E10 with this new filter and this new model. According to another embodiment, the computer implements steps E4 to E10 a predetermined number of times and selects the filter pair F / model M to obtain the lowest average error for determining the relevant point.

During an eleventh step E1 1, the computer 1 1 stores in a profile the filter F and the model M selected in the tenth step E10 and the binary code C determined in the seventh step E7. The profile may also include the first predetermined threshold used in the sixth step E6 for the determination of local extrema. This profile is recorded in the first storage means 15. When the learning curves are multi-dimensional curves, the preceding steps are applied according to each of the dimensions. According to a first variant, a profile is stored at the end of the eleventh step E1 1 for each of the dimensions. According to a second variant, only a profile including the filter / model pair of the dimension having the lowest determination error is stored at the end of the eleventh step E1 1. In these two variants each profile then integrates an indication of the dimension to which it relates. According to a third variant, the selected model M is a variable selection model satisfying the equation t = AX, where X is a vector whose components are abscissae of characteristic points according to each of the dimensions of the learning curves. Such a model is said to be multidimensional. Such characteristic points are then determined in the sixth step E6 for each of the dimensions of each learning curve. Similarly, the recurring characteristic points and the binary code are then determined during the seventh step E7 for each of these dimensions. These binary codes are also recorded in a multidimensional profile during the eleventh step E1 1, associated with an indication of the dimension to which they relate.

As represented in FIG. 8, the method of estimation on a curve TJ of a point that is relevant for the detection of anomaly of the engine 9, using profiles determined according to the learning process described above, can be implemented by the computer 1 1, according to steps F1 to F9.

This curve TJ is obtained from measurements of engine operating parameters acquired at different times by at least one sensor 10. In a first step F1, the computer selects a profile from the profiles generated by the learning process described above and stored in the first storage means 15. In a second step F2, the computer 1 1 applies at the curve TJ the filter F associated with the profile selected at the first step F1 and obtains a filtered curve.

During a third step F3, the computer 1 1 determines the characteristic points of the filtered curve obtained in the second step F2. The determination of the local extrema of the curve TJ can use the first predetermined threshold associated with the profile selected in the first step F1. From these characteristic points, the computer then determines the binary code C, each component of which codes the direction of variation of two consecutive characteristic points. Said code is determined in the same manner as the binary code determined in the seventh step E7 for the recurring characteristic points of a learning curve.

In a fourth step F4, the calculator determines whether the code C obtained in the third step F3 is identical to the code C associated with the profile selected in the first step F1.

If necessary, the shape of the curve TJ corresponds to the selected curve profile and the computer then carries out the fifth step F5 during which the selected profile is used to determine a relevant point on the curve TJ. In the opposite case, the curve TJ does not correspond to the selected profile and the calculator 11 again implements the steps F1 to F4.

During the fifth step F5, the calculator determines a relevant point, for example its abscissa, on the curve TJ from the points characteristics determined in the third step F3 and the model M associated with the profile selected in the first step F1.

The TJ curve can also be multidimensional. According to a first variant, if the profiles memorized during the learning phase are all relative to one and the same dimension, steps F1 to F5 above are applied to this dimension. According to a second variant, if profiles each associated with one dimension have been memorized during the learning phase for at least two of the dimensions of the curve V, steps F1 to F5 are applied separately to each of these dimensions and a point The relevant average is determined from the relevant points determined according to each of the dimensions, the abscissa of the average relevant point being able for example to be an average of the abscissas of the relevant points determined according to each of the dimensions. According to a third variant, if the stored profiles are multidimensional, each comprising a multidimensional model, the steps F1 to F5 described above are then applied so that, during the third step F3, the characteristic points of the curve and a code binary are determined for each dimension of the curve. The computer then determines in the fourth step F4 the most appropriate multidimensional profile to the curve from these binary codes and binary codes recorded in the selected multidimensional profile.

During a sixth step F6, the computer 1 1 can use one or more relevant points determined by the implementation of steps F1 to F5 to estimate at least one specific indicator representative of the operating state of the engine 9. As indicated previously, such indicators can be the duration of the different engine start phases, the ignition time, the stopping time or the maximum and average gradients of the exhaust gas temperature. Different treatments can be implemented from these indicators.

A first treatment may consist of a diagnosis of the state of the engine at the moment of acquisition of the curves used to determine said indicators. During a seventh step F7, the computer thus uses the indicators to estimate whether the engine has a malfunction that could justify a return to the workshop for maintenance, for example to replace a defective part.

A second treatment may consist of a prognosis of a future degradation of the operation of the engine from successive measurements. During an eighth step F8, the indicators determined from measurements relating to a flight of the engine are thus stored and this step is repeated flight after flight in order to obtain a succession of indicators whose evolution over time is representative of the evolution of the operating state of the engine 9. During a ninth step F9, the computer then implements a prognostic process of a future deterioration of the state of the engine from the evolution over time indicators memorized flight after flight at the eighth step F8.

Claims

1. Method for estimating on a curve a relevant point for detecting an anomaly of an engine (9), said curve representing an evolution as a function of time of physical operating parameters of the engine measured by at least one sensor (10 ) on said engine,

implemented by a computer (1 1) connected to first storage means (15),

said first storage means (15) memorizing at least one profile comprising a binary code each component of which encodes a direction of variation between two consecutive characteristic points of at least one learning curve, a model making it possible to estimate a relevant point at from a set of characteristic points of a curve, and a filter, said method comprising:

a/ (F1) the selection of a profile stored in the first storage means (15);

bl (F2) applying the filter of the selected profile to said curve; c/ (F3) determining a set of characteristic points of said filtered curve and a binary code each component of which encodes the direction of variation of two consecutive characteristic points belonging to said set of characteristic points;

61 (F4) comparison of the determined code and the code of the selected profile;

e/ (F5) depending on said comparison, the estimation of the relevant point on said curve from the characteristic points of said filtered curve and the model of the selected profile;

2. Method according to claim 1, in which if the determined code is different from the code of the selected profile, a new profile stored in the first storage means is selected and the computer executes steps b1 to e/ again;

3. Method according to one of claims 1 to 2, in which the relevant point is chosen from an instant of opening of a valve, an instant of net variation of a temperature or a pressure, an instant of reaching a certain speed by a high pressure compressor or a low pressure compressor, an instant of disengagement of a starter;

4. Method according to one of claims 1 to 3, in which the characteristic points of curves are chosen from inflection points, local extrema, abrupt changes in slopes;

5. Method according to one of claims 1 to 4, in which a profile further comprises a threshold and in which the characteristic points are consecutive local extrema whose ordinate difference is greater than said threshold;

6. Method according to one of claims 1 to 5, in which said models are generalized linear models with selection of variables;

7. Method according to one of claims 1 to 6, in which said models verify the equation: t = AX

where - 1 is the abscissa of the relevant point to estimate,

- A is a line vector containing regression coefficients,

- X is a column vector whose components are abscissa of the characteristic points and their transforms;

8. Method according to one of claims 1 to 7, further comprising (F6) a step of estimation from the estimated relevant points of specific indicators chosen for their representativeness of the operating state of the engine;

9. Method according to claim 8, further comprising (F7) a step of diagnosing the engine based on the estimated specific indicators;

10. Method according to claim 8, comprising (F8) a step of memorizing in storage means, flight after flight, estimated specific indicators and (F9) a step of prognosticating a deterioration in the operating state of the engine from the evolution of the specific indicators stored;

1 1 . Method according to one of claims 1 to 10, in which each profile stored in the first storage means is determined by a learning process;

12. Method according to claim 1 1, in which the learning process for a profile comprises:

a/ (E1) the display by a display device (19) of several learning curves;

b/ (E2) the determination by said calculator (1 1) of a relevant point for the detection of anomaly on each of the learning curves, said relevant point of each of the learning curves being selected by an expert at l help of input means (20);

cl (E3) storing in second storage means (21) each of said learning curves associated with said determined relevant point;

d/ (E4) the selection by said calculator (1 1) of a filter and a model;

e/ (E5) the application by said calculator (1 1) of the selected filter to each of the learning curves;

f/ (E6) the determination by said calculator (1 1) of the characteristic points of each of the filtered learning curves; g/ (E7) the determination by said calculator (1 1) among the determined characteristic points of recurring characteristic points appearing on each filtered learning curve and of a binary code each component of which codes the direction of variation between two recurring characteristic points consecutive;

h/ (E8) from the recurring characteristic points determined and the model selected, the estimation by said calculator (1 1) of the relevant point;

M (E9) the estimation by said calculator (1 1) of the error associated with the filter and model selected in step d/ by comparison of the relevant point estimated with the relevant point selected by the expert in step b / ;

\l (E1 1) the storage, in said first storage means (15), of a profile comprising said determined binary code, said filter and said model making it possible to minimize the estimated error or to obtain an estimated error less than a determined threshold (E10);

13. Computer program product comprising program code instructions for executing the steps of the method according to any one of the preceding claims when said program is executed on a computer;

14. Data processing system comprising a calculator (1 1), input means (20), at least one display device (19) characterized in that it is configured to implement the steps of the method according to any one of claims 1 to 12.